Google Cloud announced on the 21st local time that it plans to expand infrastructure based on NVIDIA's Vera Rubin platform to multiple computing sites, with up to 960,000 Rubin GPUs deployable. As the battle for AI computing power reaches a fever pitch, this figure pushes next-generation accelerator scale directly into the millions, as Google seeks to gain a head start in the cloud market arms race.
In step with the expansion, Google Cloud's A5X instances based on NVIDIA's Vera Rubin NVL72 rack-scale systems have been released. These instances use NVIDIA ConnectX-9 SuperNICs, paired with Google Cloud's Virgo network and fourth-generation liquid cooling technology, and are set to launch next month. From rack-scale systems to liquid cooling solutions, Google is paving the way for high-density, long-duration workloads—with heat dissipation at a single data center approaching its limits, liquid cooling has become a must-answer question for hosting massive GPU fleets.
Confidential computing arrives: even bare metal can guard against insiders
Beyond computing power, Google Cloud is also playing a security card: it will become one of the first cloud providers to offer rack-scale NVIDIA confidential computing services on bare metal instances; in addition, Google Cloud will also officially launch the service on G4 virtual machines equipped with NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs.
The core of confidential computing is keeping data encrypted and isolated even during processing, so that even those holding underlying privileges find it hard to peek. For sensitive industries such as finance, healthcare, and government, this is precisely the prerequisite for moving from being afraid to move to the cloud to daring to do so. By covering both rack-scale bare metal and general-purpose virtual machines, Google is playing both sides between "strongest computing power" and "most solid protection."
Looking at the whole chessboard, Google's moves reflect a shared anxiety among cloud providers: the stronger large models become, the more bottomless the thirst for underlying computing power. Amazon, Microsoft, and Meta are likewise weighing their own chips against NVIDIA supply, while Google is betting on both its in-house TPUs and Vera Rubin, following a "dual-track supply security" route. As 960,000 Rubin GPUs move from blueprint to server rooms, the ceiling of AI infrastructure is being raised to a new height.